Skip to content
agent-bridge

Dashboard

The dashboard starts automatically: whichever agent session hosts the bridge also hosts the dashboard at http://127.0.0.1:4777, and if that session ends, another one takes it over. To open it:

  • /agent-bridge:dashboard in Claude Code (or ask any agent to “open the agent-bridge dashboard”), or

  • from any terminal:

    Terminal window
    npx -y github:rennerdo30/agent-bridge ui     # opens the running dashboard, or starts one if no agent runs

Turn the automatic start off with "dashboard": false in ~/.agent-bridge/config.json (or AGENT_BRIDGE_DASHBOARD=off); change the port with "dashboardPort".

The dashboard has a sessions sidebar, grouped by PC, with search and folding:

  • Overview: connected sessions, usage left, working and finished jobs, and message history with a box to send a message yourself (as “you”). Finished jobs show merge/review outcomes.
  • Session page: its own chat with an owner-message composer for local sessions, and a subagent tree with nested bridge jobs and the CLIs’ own subagents. Native children offer supported direct input or a note to their parent. Conversations use chat bubbles; commands fold into rows, and follow-ups stay in the same job. Progress and ETA appear in the job row and conversation header. Older finished runs fold into an archive and remain readable.
  • Search history: search messages, subagent runs, Decisions and CLI chats, including archived history; filter by kind or agent, inspect sources, or opt into a model-generated summary.
  • Decisions: pinned owner choices with scope and revision history.
  • Waiting for you: owner questions with recommended options and a free-text answer composer, alongside pending tool approvals with countdowns and allow/deny controls. Question sound, desktop toasts and reminders are configurable. Answers remain separate from tool permission.
  • Network: pairing, discovery details, connection checks, firewall guidance and file-transfer progress. Paired sessions appear in the sidebar, with their chats and subagents readable over the authenticated link.

These screenshots use synthetic demo data only.

Overview · dark Overview · light
Dashboard overview in dark theme Dashboard overview in light theme
Session · dark Session · light
Demo session chat, nested bridge jobs, project main and composers in dark theme Demo session chat, nested bridge jobs, project main and composers in light theme
Network Search history
Network page with a connected demo PC Search history with synthetic settings results
A demo owner question in Waiting for you with recommended options and an answer composer

It only listens on 127.0.0.1. Its link contains an owner-only launch key stored in ~/.agent-bridge/dashboard-key; opening that link establishes the dashboard’s HttpOnly cookie. The key survives updates and restarts. Requests need authentication and local Host checks; writes also require the dashboard header. See security. ui options: --port=N, --no-open.

Local sessions also have read-only transcript APIs for their normal chat and their CLI’s native subagents: Claude Code JSONL, Codex rollout JSONL, and OpenCode SQLite. They use the peer’s session id, the same dashboard cookie, and bounded incremental reads. CLI files are never edited. CLAUDE_CONFIG_DIR, CODEX_HOME, and OpenCode’s XDG_DATA_HOME storage override are respected. Paired-PC reads use the existing TLS link; update and restart hosting sessions on both PCs. See the transcript API and paired dashboard reads.

Run logs

Section titled “Run logs”

Every run also writes a step-by-step log to ~/.agent-bridge/runs/, whose path is in the result. Follow a run live from any terminal:

Terminal window
npx -y github:rennerdo30/agent-bridge watch            # the newest run
npx -y github:rennerdo30/agent-bridge watch opencode   # the newest opencode run

opencode subagents keep opencode’s full tool set on purpose, because opencode’s free tier rejects subagents with a restricted tool list. Their prompt tells them to only relay.